Re: distributed filesystem for backup

[Matt Cuttler <mcuttler@bnl.gov>]

Jorge Salamero Sanz wrote:
>
>
> gfs does need a dedicated network ?
>   

Probably not an absolute/strict requirement (at the very least highly
recommended).

If you really only have one eth interface to use, you might be able to
get away with using IP aliases. Probably better if in a completely
different private IP range. This should work if all of your cluster
nodes are off of the same layer 2 switch or daisy-chained layer 2
switches (i.e. all bets are off if you or your network engineers are
doing anything with layer 3 VLANs, or otherwise limiting your ability to
use additional IP address ranges on your network ports).

I once did something similar with OCFS; it "required" dedicated network
interfaces for the internode communication.


At any rate, I'm sure you've thought about all the reasons you should
use a dedicated network for GFS. But if you haven't :-)

You mention that your cluster is used for MPI. If you run GFS off of the
same network interfaces, you're (obviously) going to limit your node's
available bandwidth. There's the additional GNBD traffic (since you're
not using fibre channel), the DLM traffic etc. Of course, this may not
matter to you, if you're using this filesystem for occasional backups
and such.

A final thought: You might want to check out PVFS/PVFS2. Your cluster
users might appreciate it, too since it has MPI-IO support, and solves
the old "NFS/MPI problem".

-Matt Cuttler